🧬 biology

$p$ -adic Bi-Filtrations for Topological Machine Learning on Genomic Sequences

The paper introduces pVR, a novel topological machine learning framework that leverages $p$ -adic numbers and bi-filtered Vietoris–Rips complexes to achieve superior alignment-free genomic sequence classification, particularly in low-sample regimes, by jointly encoding hierarchical positional structure and local compositional content.

Original authors: Tirtharaj Dash, Gunja Sachdeva

Published 2026-06-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tirtharaj Dash, Gunja Sachdeva

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to sort a massive library of DNA books. Each book is a long string of letters (A, C, G, T) representing an organism's genetic code. Your goal is to figure out which books belong to the same "family" (species, virus variant, etc.) without reading every single word in every book.

This paper introduces a new tool called pVR to solve this sorting problem. It combines two very different ways of looking at these genetic books to create a smarter classification system, especially when you don't have many books to work with.

Here is how it works, broken down into simple concepts:

1. The Two Ways to Look at a Book

Most traditional methods look at DNA in just one way: they count how often certain short phrases (like "ATG" or "CGT") appear. It's like judging a book only by its vocabulary. If two books use the same words, they are considered similar.

The pVR tool, however, looks at the DNA in two complementary ways simultaneously:

The "Tree" View (p-adic distance): Imagine the DNA sequence as a family tree. The first few letters are the "grandparents," the next few are the "parents," and the end letters are the "children." This method cares deeply about the order and position of the letters. A change at the very beginning of the sequence is treated as a massive family split, while a change at the very end is a minor detail. It captures the hierarchical structure of the sequence.
The "Recipe" View (Compositional distance): This is the traditional method. It ignores the order and just asks, "What ingredients are in this soup?" It counts the total amount of A's, C's, G's, and T's. It captures the local content but ignores the structure.

2. The Problem with Looking at Just One View

The paper proves a surprising mathematical fact: if you only use the "Tree" view (the p-adic method), the math says you can't actually see any interesting shapes or patterns. It's like looking at a forest from directly above; you just see a flat map of trees, but you miss the loops, tunnels, and 3D structures that exist between them.

However, if you only use the "Recipe" view, you miss the deep evolutionary family connections.

3. The Solution: The "Bi-Filtration" Net

To fix this, pVR builds a double-layered net (called a bi-filtered complex).

Imagine you are trying to find groups of friends in a crowded room.
Layer 1: You only let people hold hands if they share the same last name (the "Tree" view).
Layer 2: You only let people hold hands if they are wearing the same color shirt (the "Recipe" view).

pVR only connects two DNA sequences if they pass both tests at the same time. By combining these two rules, the tool reveals hidden shapes (loops and holes) in the data that neither rule could find alone. These shapes act as a unique "fingerprint" for the group of DNA sequences.

4. How It Performs (The Results)

The researchers tested this tool on 12 different real-world datasets, ranging from human viruses to animal mitochondria.

When Data is Scarce (The "Small Library"): When there are very few DNA sequences to analyze (like a new, rare virus), pVR is a superstar. It outperformed four other standard methods on three out of six small datasets. In one case (Ebola virus classification), it was 21 percentage points more accurate than the next best method. It even beat a massive, 500-million-parameter AI model (Nucleotide Transformer) on these small tasks.
- Why? Because pVR uses the "Tree" structure as a smart guess (a prior) about how life evolves, which helps when there isn't enough data to learn from scratch.
When Data is Abundant (The "Big Library"): When there are thousands of sequences, all methods (including simple ones) become very accurate. pVR stays competitive but doesn't win big. This is because when you have enough data, the simple "Recipe" view is usually enough to get the job done.
When It Struggles: The tool did worse on a specific SARS-CoV-2 dataset. The authors explain this is because those virus variants didn't evolve in a neat "tree" pattern; they changed via scattered, random mutations. Since pVR relies on that tree-like structure, it got confused.

5. The Bottom Line

The paper claims that pVR is a specialized tool for "small data" problems in genomics.

Analogy: If you have a huge library, a simple index card (counting words) is enough to find a book. But if you only have a few books and they are very similar, you need a tool that understands the family history and the story structure to tell them apart. pVR is that tool.
Key Takeaway: It works by mathematically proving that combining a "hierarchical" view (family tree) with a "compositional" view (ingredient list) creates a powerful new way to classify DNA, especially when you don't have much data to work with.

The code for this tool is publicly available, and the authors emphasize that it is fast (running in seconds on a standard computer) and robust, meaning the results don't change just because you tweak the mathematical settings slightly.

Technical Summary: p-adic Bi-Filtrations for Topological Machine Learning on Genomic Sequences

Problem Statement

Comparative genomic analysis requires classifying organisms, identifying variants, and reconstructing evolutionary relationships. While classical alignment-based methods (e.g., MAFFT, MUSCLE) are effective for closely related sequences, they scale poorly with length and divergence. Alignment-free methods (e.g., k-mer frequency, Natural Vector, MinHash) address scalability but typically treat k-mers as independent features, discarding the structural and hierarchical relationships inherent in genetic codes. Furthermore, recent topological data analysis (TDA) approaches in genomics have largely relied on frequency-based or sketch-based distances, potentially missing the hierarchical arithmetic structure of biological sequences. Conversely, p-adic models have been used to encode biological sequences based on ultrametric properties, yet they have not been combined with simplicial persistent homology for classification tasks.

Methodology: The pVR Framework

The authors introduce pVR, an alignment-free framework that integrates p-adic number theory with topological data analysis. The core innovation is a bi-filtered Vietoris–Rips (VR) complex parameterized by two complementary distance metrics derived from each DNA sequence:

Hierarchical Axis (p-adic Distance, $D_p$ ):
- Nucleotides are mapped to digits in $\mathbb{Z}_p$ (specifically $p=5$ for DNA).
- k-mers are encoded as p-adic integers, where the first nucleotide corresponds to the lowest-order digit.
- The distance between k-mers is defined by the p-adic metric, where sequences sharing a long common prefix are "close."
- At the sequence level, $D_p$ is a weighted $L_1$ distance between multi-scale histograms of p-adic residues, capturing hierarchical prefix structure.
Compositional Axis (Compositional Distance, $D_c$ ):
- This is the standard $L_1$ distance between normalized k-mer frequency vectors.
- It captures local sequence content without regard to position or hierarchy.

The Bi-Filtration:
Instead of applying persistent homology to a single metric space, pVR constructs a bi-filtered VR complex $VR(\epsilon_p, \epsilon_c)$ . A simplex exists in this complex only if the pairwise distance between its vertices satisfies both thresholds: $d_p(x,y) \le \epsilon_p$ and $d_c(x,y) \le \epsilon_c$ .

Feature Extraction:
From the bi-filtration grid, the framework extracts:

Degree Profiles: The number of neighbors for each sequence vertex at each grid point, capturing local connectivity evolution.
Global Betti Numbers: $\beta_0$ (connected components) and $\beta_1$ (loops) for the entire complex at each threshold.
Static Features: Concatenated p-adic histograms and k-mer frequency vectors.
These features are fed into standard classifiers (XGBoost, SVM, 5-NN).

Key Contributions

Theoretical Contributions

The paper establishes four structural results:

Triviality of Single-Axis Ultrametric Topology: The authors prove (Theorem 5) that Vietoris–Rips complexes built from finite strictly ultrametric spaces have trivial homology in all positive dimensions ( $H_k = 0$ for $k \ge 1$ ). This explains why a single p-adic axis alone cannot capture higher-order topological features (loops).
Nontriviality of Bi-Filtration: They prove (Proposition 8) that pairing a strictly ultrametric distance with a compositional metric can recover nontrivial homology (specifically 1-cycles) that is invisible to either filtration alone.
Stability: The resulting bi-persistence module is proven to be stable under metric perturbations (Proposition 11), ensuring robustness to noise.
Prime Invariance: The features are invariant to the choice of the prime $p$ (provided $p \ge |\Sigma| + 1$ ), eliminating the prime selection as a hyperparameter (Proposition 12).

Note: The authors clarify that while the theoretical proofs assume strict ultrametricity, the implemented sequence-level distance $D_p$ is only approximately ultrametric. However, empirical results confirm the predicted behavior where single-axis p-adic filtration yields trivial topology, while the bi-filtration recovers structure.

Empirical Contributions

Performance on Low-Sample Data: On six low-sample benchmarks (28–150 sequences), pVR outperforms four established alignment-free baselines (FFP-JS, NVM, MinHash, k-mer frequency) on three datasets (Ebola, Influenza HA, Mammalian mitochondrial).
- Ebola: pVR-SVM achieved 100.0% accuracy, a 21.1 percentage point gain over the best baseline (MinHash at 78.9%).
- Influenza HA: pVR achieved 72.5% vs. 62.2% for FFP-JS.
- Mammalian: pVR achieved 62.0% vs. 53.9% for FFP-JS.
Comparison with Foundation Models: On three low-sample benchmarks, pVR outperformed zero-shot embeddings from the 500M-parameter Nucleotide Transformer v2 by 6.7 to 11.4 percentage points.
Large-Sample Regime: On six large-sample datasets (up to 500 sequences), performance across all methods saturated (mostly 96–100%), with pVR remaining competitive but not significantly outperforming simple compositional baselines.
Failure Case: pVR underperformed on the SARS-CoV-2 variant benchmark (small sample). The authors attribute this to the fact that SARS-CoV-2 variants diverge via scattered point mutations rather than deep hierarchical clade divergence, violating the structural assumption of the p-adic axis.

Significance and Claims

The paper claims that pVR provides a mathematically grounded method for incorporating hierarchical structural priors into genomic machine learning. Its primary significance lies in demonstrating that:

Structural Priors Matter: In low-data regimes, encoding the hierarchical nature of evolutionary divergence (via p-adic numbers) combined with compositional content yields superior inductive bias compared to purely frequency-based methods or large pre-trained models.
Topological Necessity of Bi-Filtration: The work provides theoretical justification for why combining an ultrametric distance with a non-ultrametric one is necessary to extract nontrivial topological features from biological data.
Efficiency: The method is computationally efficient, processing datasets with up to 500 sequences in under 30 seconds on a standard workstation without GPU acceleration.

The authors modestly conclude that the method is most effective when data is scarce and divergence is hierarchical (e.g., emerging pathogens, rare species). They acknowledge limitations, noting that the method struggles when divergence is driven by point mutations (SARS-CoV-2) or recombination (HEV), and that the $O(N^3)$ complexity of the bi-filtration grid loop becomes a bottleneck for very large datasets ( $N > 500$ ).